Re: route cache DoS testing and softirqs

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

On Tue, Mar 30, 2004 at 01:07:12AM +0000, Andrea Arcangeli wrote:
> btw, the set_current_state(TASK_INTERRUPTIBLE) before
> kthread_should_stop seems overkill w.r.t. smp locking, plus the code is
> written in the wrong way around, all set_current_state are in the wrong
> place. It's harmless but I cleaned up that bit as well.

I think set_current_state(TASK_INTERRUPTIBLE) before kthread_should_stop()
_is_ required, otherwise kthread_stop can fail to destroy a kthread.


kthread_stop does:

	1. kthread_stop_info.k = k;
        2. wake_up_process(k);

and if ksoftirqd were to do :

	a. while (!kthread_should_stop()) {
        b.         __set_current_state(TASK_INTERRUPTIBLE);
        c.         schedule();
           }


There is a (narrow) possibility here that a) happens _after_ 1) as well as 
b) _after_ 2).

In such a case kthread_stop would have failed to wake up the kthread!


The same race is avoided if ksoftirqd does:

        a. __set_current_state(TASK_INTERRUPTIBLE);
	b. while (!kthread_should_stop()) {
        c.         schedule();
        d.         __set_current_state(TASK_INTERRUPTIBLE);
           }

        e. __set_current_state(TASK_RUNNING);

In this case, even if b) happens _after_ 1) and c) _after_ 2), schedule
simply returns immediately because task's state would have been set to 
TASK_RUNNING by 2). It goes back to the kthread_should_stop() check and 
exits!








-- 


Thanks and Regards,
Srivatsa Vaddagiri,
Linux Technology Center,
IBM Software Labs,
Bangalore, INDIA - 560017